op-conductor: fixes flashblocks handler leaks

[zhwrd]

Fixes a memory leak in op-conductor flashblocks websocket handler:

• Broadcasts now drop and close slow clients when their send queue is full, preventing large buffered flashblock payloads from piling up.

Implemented some additional flashblock memory pressure and shutdown leakage (not an issue in production generally since the container will get killed, but included for completeness)

• Hub is now protected with a mutex, waits for a stopped signal on shutdown, and the stop path drains clients safely without leaving goroutines blocked on unregister. The read pump now unregisters directly when the hub is already stopping, so stopping won’t strand clients or buffers.
• Flashblocks handler Stop now waits for the hub to finish and guards the done-channel close. Also, the rollup-boost read loop idles when not leader to avoid needless allocations while a follower.

https://github.com/ethereum-optimism/k8s/issues/8053